A resume was never a measurement of capability. It was always a proxy, a compressed, self-reported summary that a hiring team used because getting the real signal, watching someone actually do the work, was expensive and slow. Proxies are fine right up until something breaks the correlation between the proxy and the thing it's standing in for. In AI hiring specifically, that correlation just broke. The tools that make it trivially easy to polish, embellish, and pad a resume are the same tools whose absence in a candidate's actual skill set is exactly what you're trying to screen for. Relying on the resume now doesn't just fail to help, it actively misleads.
A resume was always a stand-in, never the thing itself
Nobody has ever actually wanted to know what's on a candidate's resume. What every hiring manager actually wants to know is whether this specific person can do this specific job well. The resume exists because that direct question is expensive to answer, so companies settled for a cheaper, faster proxy: a self-authored summary of past titles, tools, and achievements. That proxy worked reasonably well for a long time because writing a convincing, accurate resume correlated loosely with having done the work it described. It was never a great signal, ask any hiring manager about a resume that oversold someone badly, but it was good enough to be worth the shortcut.
Why the correlation just broke, specifically in AI hiring
Two things happened at once, and they compound. First, AI writing tools made it trivial to take a thin set of real accomplishments and produce a polished, quantified, achievement-dense resume that reads as strong regardless of how strong the underlying work actually was. The skill being screened for, and the skill now doing the screening's dirty work, are uncomfortably close to the same skill. Second, credential inflation in AI specifically is severe: 'led our AI initiative,' 'built our RAG pipeline,' 'shipped LLM features' now appear on resumes describing everything from a genuinely hard production system to a weekend prototype that never left a notebook. The title costs the same number of words either way. A resume that would have reliably signaled real experience three years ago now signals almost nothing about whether the underlying system worked, scaled, or shipped.
- AI writing assistance closes the gap between a thin accomplishment and a polished description of it, at scale, for free.
- 'AI experience' as a phrase now covers everything from production-grade systems to a single class project, with no way to tell which from the resume alone.
- Quantified bullet points ('improved accuracy by 40%') are trivially easy to generate and nearly impossible to verify from the document itself.
- The candidates most disadvantaged by this are often the strongest builders, people who'd rather spend the evening shipping than polishing prose.
What a resume can't show, no matter how well it's written
The actual differentiator in AI hiring is whether someone can take an ambiguous problem, real data, real constraints, a real deadline, and ship a working system, then debug it when it breaks in production in a way no test set predicted. That capability lives in judgment under constraint: what tradeoffs someone makes when the model is slow, the data is messy, and the deadline hasn't moved. None of that is expressible in a bullet list, because a bullet list describes outcomes, not the decisions that produced them, and outcomes can be described accurately by people who didn't make the hard calls themselves, or embellished by people who did.
What's actually replacing the resume
| Signal | What it verifies that a resume can't |
|---|---|
| Real work samples (shipped code, a live system, a public repo) | The work exists, is inspectable, and reflects actual decisions, not a description of decisions |
| Structured technical evaluation against a fixed rubric | Consistent, comparable evidence of how someone reasons through an ambiguous, realistic problem live |
| Reference checks that ask for specifics, not ratings | What this person actually did under pressure, described by someone who watched it happen |
| A short paid trial or scoped project | Whether the work holds up outside an interview room, under real constraints and real feedback |
The reference check questions that actually work
Most reference checks ask 'would you rehire them?' and 'how would you rate their communication?' and get back polite, low-information answers, because the format invites politeness, not detail. The reference checks that actually predict future performance ask for a specific story: 'Tell me about a time this person's system broke in production. What did they do?' or 'Describe the hardest technical tradeoff you watched them make.' Specific questions produce specific, checkable answers, and a reference who can't produce one, who stalls on specifics and retreats to generic praise, is itself informative.
This isn't an argument against screening, it's an argument against the wrong screen
None of this means screening should get looser, if anything it should get stricter. It means the object being screened needs to change from a self-reported document to something that can't be talked into looking better than it is: real code, a structured live evaluation, a reference asked the right question. Companies that keep leaning on resume screening as the primary filter aren't being rigorous, they're optimizing for who writes the most convincing document, which is a real skill, just not the one the role actually requires.